CASCADE + BERT: Using Context Embeddings and Transformers to Predict Sarcasm

Sarcasm is a form of verbal irony, in which a writer or speaker states the opposite of their intended message in order to mock or show contempt. Sarcasm is commonly used online, and being able to detect sarcasm is crucial to understand and classify online pieces of text. In this work, we attempt to classify Reddit comments (from the SARC corpus) as either sarcastic or sincere. Sarcasm is heavily reliant on context -- information about the world or a text author. Therefore, we propose a BERT model, augmented with three different context types: discourse context, user context, and community context. After testing this model, we found that the addition of user context shows increased performance compared to training without user context. However, incorporating community context did not improve performance. We were able to achieve a maximum accuracy of 75.8 percent.